AMD Launches Helios to Challenge Nvidia's Grip on AI Infrastructure

On July 22, 2026, AMD held its "Advancing AI" conference in San Francisco and officially launched Helios, a rack-scale AI system designed to challenge Nvidia's grip on the data-center market. A single Helios rack combines 72 new Instinct GPUs, 18 EPYC server CPUs and 31 terabytes of high-bandwidth memory, all cooled by liquid. AMD says the design is built to train and run the largest frontier AI models at scale.

According to AMD, Helios delivers about 15 percent more raw AI compute, 50 percent more memory capacity and up to 30 percent more tokens per dollar than Nvidia's current Vera Rubin rack. The company is also pushing a new open rack standard, so customers can mix Helios-style hardware with parts from other vendors instead of being locked into a single supplier.

Big buyers are already lining up. Microsoft confirmed it will deploy Helios inside its Azure cloud, while OpenAI, Meta, Oracle and Anthropic have all agreed to use the platform at gigawatt scale. As part of the deal, Anthropic will host up to two gigawatts of GPU capacity on Helios, and OpenAI took a small equity stake in AMD to lock in long-term supply.

The launch matters beyond AMD itself. The CEO raised her forecast for the total AI accelerator market from earlier estimates to roughly 1.4 trillion dollars by 2030, almost the size of today's entire semiconductor industry. For years, the AI build-out depended on one company's roadmap; Helios is the clearest sign yet that the supply chain is starting to open up.